On Tue, Feb 4, 2025 at 9:25 AM Alan Maguire <alan.maguire@xxxxxxxxxx> wrote:
>
> On an aarch64 kernel with CONFIG_PAGE_SIZE_64KB=y (64k pages),
> arena_htab tests cause a segmentation fault and soft lockup.
>
> $ sudo ./test_progs -t arena_htab
> Caught signal #11!
> Stack trace:
> ./test_progs(crash_handler+0x1c)[0x7bd4d8]
> linux-vdso.so.1(__kernel_rt_sigreturn+0x0)[0xffffb34a0968]
> ./test_progs[0x420f74]
> ./test_progs(htab_lookup_elem+0x3c)[0x421090]
> ./test_progs[0x421320]
> ./test_progs[0x421bb8]
> ./test_progs(test_arena_htab+0x40)[0x421c14]
> ./test_progs[0x7bda84]
> ./test_progs(main+0x65c)[0x7bf670]
> /usr/lib64/libc.so.6(+0x2caa0)[0xffffb31ecaa0]
> /usr/lib64/libc.so.6(__libc_start_main+0x98)[0xffffb31ecb78]
> ./test_progs(_start+0x30)[0x41b4f0]
>
> Message from syslogd@bpfol9aarch64 at Feb 4 08:50:09 ...
> kernel:watchdog: BUG: soft lockup - CPU#1 stuck for 26s! [kworker/u8:4:7589]
>
> The same failure is not observed with 4k pages on aarch64.
>
> Investigating further, it turns out arena_map_free() was calling
> apply_to_existing_page_range() with the address returned by
> bpf_arena_get_kern_vm_start(). If this address is not page-aligned -
> as is the case for a 64k page kernel - we wind up calling apply_to_pte_range()
> with that unaligned address. The problem is apply_to_pte_range() implicitly
> assumes that the addr passed in is page-aligned, specifically in this loop:
>
> do {
> if (create || !pte_none(ptep_get(pte))) {
> err = fn(pte++, addr, data);
> if (err)
> break;
> }
> } while (addr += PAGE_SIZE, addr != end);
>
> If addr is _not_ page-aligned, it will never equal end exactly.
>
> One solution is to round up the address returned by bpf_arena_get_kern_vm_start()
> to a page-aligned value. With that change in place the test passes:
>
> $ sudo ./test_progs -t arena_htab
> Summary: 1/1 PASSED, 1 SKIPPED, 0 FAILED
>
> Reported-by: Colm Harrington <colm.harrington@xxxxxxxxxx>
> Signed-off-by: Alan Maguire <alan.maguire@xxxxxxxxxx>
> ---
> kernel/bpf/arena.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/kernel/bpf/arena.c b/kernel/bpf/arena.c
> index 870aeb51d70a..07395c55833e 100644
> --- a/kernel/bpf/arena.c
> +++ b/kernel/bpf/arena.c
> @@ -54,7 +54,7 @@ struct bpf_arena {
>
> u64 bpf_arena_get_kern_vm_start(struct bpf_arena *arena)
> {
> - return arena ? (u64) (long) arena->kern_vm->addr + GUARD_SZ / 2 : 0;
> + return arena ? (u64) round_up((long) arena->kern_vm->addr + GUARD_SZ / 2, PAGE_SIZE) : 0;
Thanks for the report. The fix is incorrect though.
GUARD_SZ/2 is 32k,
so with roundup the upper guard is gone.
We probably need to:
-#define GUARD_SZ (1ull << sizeof_field(struct bpf_insn, off) * 8)
+#define GUARD_SZ round_up(1ull << sizeof_field(struct bpf_insn, off)
* 8, PAGE_SIZE << 1)
Better ideas?
pw-bot: cr
